Using Traditional and Deep Machine Learning to Predict Emergency Room Triage Levels
Mehmet Yıldırım1, Savaş Sezik2, Ayşe Başar3
1Computer Engineering Department, Boğaziçi University, Istanbul, Türkiye.
Summary
Predicting emergency room triage levels is improved by machine learning models. Combining patient history with text embeddings like Word2Vec and BERT enhances accuracy for better patient care.
Area of Science:
- Medical Informatics
- Natural Language Processing
- Machine Learning
Background:
- Accurate emergency room (ER) triage is vital for patient care and resource management.
- Developing predictive models for ER triage presents challenges, especially with non-English languages and specific healthcare systems.
- Existing methods often lack robust handling of unstructured clinical notes.
Purpose of the Study:
- To develop and evaluate machine learning models for predicting ER triage levels using Turkish patient data.
- To compare the effectiveness of traditional machine learning algorithms against deep learning approaches for this task.
- To assess the impact of various text embedding techniques (Bag of Words, Word2Vec, BERT) on model performance.
Main Methods:
- Utilized traditional machine learning (logistic regression, random forest, XGBoost) and deep learning (CNN, LSTM, attention mechanisms) models.
- Processed unstructured Turkish patient complaints using Bag of Words, Word2Vec, and BERT embeddings.
- Incorporated structured data, including patient history and diagnosis, as features in the predictive models.
- Evaluated models on a dataset from a Turkish hospital's emergency department visits.
Main Results:
- BERT embeddings significantly improved neural network model performance.
- Word2Vec embeddings yielded better results with traditional machine learning models.
- The optimal model, XGBoost with Word2Vec embeddings, achieved 86.7% AUC, 81.5% accuracy, and 68.7% weighted F1 score.
- Integration of patient history and text embeddings substantially boosted predictive accuracy.
Conclusions:
- Text embedding and machine learning methods are effective for predicting ER triage levels.
- The XGBoost model with Word2Vec embeddings demonstrated superior performance.
- Incorporating patient history and advanced text processing techniques is crucial for enhancing triage prediction accuracy in diverse medical contexts.
